how does the inference model trained? about graphonomy HOT 4 CLOSED

1. inference model uses transfer.
2. I think the training may be better.
3. If the image resolution is so small (e.g. 64 x 64), the results may be worse. ( I don't understand (I just change scale list into 2))
4. You can use our inference model to deal with the setting of only one person in your picture. If you want to train ATR or LIP using deeplabv3, you can use the deeplabv3+ model directly.
   You may need to download the LIP dataset, change the dataloader( you can refer the cihp.py ), and then change the model and forward in the train.py.
   Maybe it can help you.

from graphonomy.

Thankyou for your advice
1 . Scale_list is the variable in inference.py in line 110.
I change the default list into ［1.25，1.5，1.75］to extract my 100*200 picture is better than default ，if retrain maybe i should change the training scale ?
but my picture(person detection box) is various size 100×200 ~300×700,I'm also confuse about how to resize..
2. If i change body parts into one label and run train_transfer_cihp.sh ,can you update pascal pretrained model.

from graphonomy.

1. Oh, I understand. If the image is 100*200, you could try the list with larger scale (e.g., [1.25,1.5,1.75,2,2.25]). You can change the training random scale in here.
2. I will update the other models as soon as possible after the ddl on 23 May TAT. So sorry about this.

from graphonomy.

thank u,best wishes

from graphonomy.